1 / 59

Reserve Ranges, Confidence Intervals and Prediction Intervals

Reserve Ranges, Confidence Intervals and Prediction Intervals. Glen Barnett, Insureware David Odell, Insureware Ben Zehnwirth, Insureware. Summary . Uncertainty and variability are distinct concepts fundamental difference between confidence interval and prediction interval.

dot
Download Presentation

Reserve Ranges, Confidence Intervals and Prediction Intervals

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Reserve Ranges, Confidence Intervals and Prediction Intervals Glen Barnett, Insureware David Odell, Insureware Ben Zehnwirth, Insureware

  2. Summary • Uncertainty and variability are distinct concepts • fundamental difference between confidence interval and prediction interval. • When finding a CI or PI, assumptions should be explicit, interpretable, testable - related to volatility in the past. • Differences between CI & PI explained via simple examples and then using loss triangles in the PTF modeling framework • loss triangle regarded as sample path from fitted probabilistic model • An identified optimal model in the PTF framework describes the trend structure and volatility about it succinctly – the "four pictures" • model predicts lognormal distributions for each cell + their correlations, conditional on explicit, interpretable assumptions related to past volatility • Immediate benefits include percentile and V@R tables for total reserve and aggregates, by calendar year and accident year.

  3. Variability and Uncertainty • different concepts; not interchangeable "Variability is a phenomenon in the physical world to be measured, analyzed and where appropriate explained. By contrast uncertainty is an aspect of knowledge." Sir David Cox

  4. Variability and uncertainty Process variability is a measure of how much the process varies about its mean – e.g. 2 (or ) Parameter uncertainty is how much uncertainty in some parameter estimate (e.g. var(µ) or s.e.(µ) ) or function of parameter estimates (say for a forecast mean – “uncertainty in the estimate”) Predictive variability is (for most models used) the sum of the process variance and parameter uncertainty ^ ^

  5. … 1 0 1 0 0 "Roulette Wheel" No. 0,1, …, 100 Mean = 50 Std Dev = 29 CI [50,50] Example: Coin vs Roulette Wheel In 95% of experiments with the wheel, observed number will be in interval [2, 97]. In 95% of experiments with the coin the number of heads will be in interval [40,60]. Coin 100 tosses fair coin (#H?) Mean = 50 Std Dev = 5 CI [50,50] Where do you need more risk capital? Introduce uncertainty into our knowledge - if coin or roulette wheel are mutilated then conclusions could be made only on the basis of observed data.

  6. "Roulette Wheel" No. 0,1, …, 100 Mean = ? Std Dev = ? CI [?,?] Example: Coin vs Roulette Wheel - similar thing with wheel (more complex) Parameter uncertainty increases width of prediction interval Coin 100 tosses Mean = ? Std Dev = ? CI [?,?] can toss coin 10 times first (5 heads –> est. mean 50) Process variability cannot be controlled but can be measured

  7. A basic forecasting problem Consider the following simple example – n observations Y1... Yn ~ N(µ,2) Yi = µ + ii ~ N(0, 2) Now want to forecast another observation... (Actually, don’t really need normality for most of the exposition, but it’s a handy starting point.) iid iid

  8. A basic forecasting problem Yn+1 = µ + n+1 Yn+1 = µ + n+1 µ known: known Yn+1 = µ + 0 Variance of the forecast is Var(µ) + Var(n+1) = 2 0 forecast of the error term ^ ^ ^ ^

  9. Next observation might lie down here, or up here. Similarly for future losses: may be high or low. The risk to your business is not simply from the uncertainty in the mean – V@R is related to the amount you will pay, not its mean. A basic forecasting problem (you're forecasting a random quantity)

  10. - Even when mean is knownexactly, still underlying process uncertainty (- with 100 tosses of a fair coin, might get 46 heads or 57 heads etc). -Except with modeling losses you design a model to describe what's going on with the data. -It doesn't really make sense to talk about a mean (or any other aspect of the distribution) in the absence of a probabilistic model. [If you don't have a distribution, what distribution is this "mean" the mean of?] A basic forecasting problem

  11. 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 0 10 20 30 40 50 60 70 80 90 100 • assumptions need to be explicit so you can test distribution is related to what's going on in the data • don’t want to use coin model if your data is actually coming from a roulette wheel! even worse if you don't get the mean right Of course, in practice we don't know the mean – we only have a sample to tell us about it.

  12. Let's make a different assumption to before • – now we don't know µ • we’ll have an estimate of the future mean based – through our model – on past values • but estimate of the future mean is not exactly the actual mean, even with a great model (random variation) • - the estimate is uncertain (can estimate the uncertainty - if the model is a good description)

  13. - So we will have an estimate of the mean and we'll also have a confidence interval for the mean. - interval designed so that if we were able to rerun history (retoss our coin, respin our roulette wheel), many times, the intervals we generate will include the unknown mean a given fraction of the time But that probability relies on the model... if the model doesn't describe the data, confidence interval is useless (won't have close to required probability coverage)

  14. Confidence Interval for Mean of Coin Model (100 tosses) with a small sample (10) 10 tosses, 5 heads. => p = ½ µ = 100p Var(µ) = 1002 Var(p) = 1002 ½ ½ /10= 15.82 95% CI for µ = 50  1.96x15.8 ~= (19,81) ^ ^ ^

  15. Let’s look at some real long-tail data. Has been inflation-adjusted and then normalized for a measure of exposure (exposures are fairly stable, so looks similar)

  16. - No trends in the accident year direction - Calendar years are sufficiently stable for our current purpose (one year is a bit low – it could be omitted if preferred, but we will keep it) Note a tendency to “clump” just below the mean – more values below mean: skewness

  17. On the log-scale this disappears, and the variance is pretty stable across years: (many other reasons to take logs) Skewness is removed – values much more symmetric about center

  18. Consider a single development - say DY 3: ML estimates of µ, : 7.190 and 0.210 (note that the MLE of is sn, with the n denominator, not the more common sn-1).

  19. We don’t know the true mean. Assuming random sample from process with constant mean, predict mean for next value as 7.190 - without some indication of its accuracy, not very helpful: 95% confidence interval for µ: (7.034, 7.345)

  20. Now we want to look at a prediction. Want to predict a random outcome where we don't know the mean (assuming variance known here, but generally doesn’t change prediction intervals a great deal) Note with coin tossing/roulette wheel experiment, almost never pay the mean (50). - don't pay mean, pay the random outcome - risk posed to the business is NOT just the uncertainty in the mean.

  21. To understand your business you need to understand the actual risk of the process, not just the risk in the estimate of the mean. Want to forecast another observation, so let’s revisit the simple model: Yn+1 = µ + n+1 Yn+1 = µ + n+1 ^ ^ ^ forecast of the error term

  22. µ UNKNOWN: So Yn+1 = µ + 0 Variance of forecast = Var (µ) + Var(n+1) = 2/n + 2 (in practice you replace 2 by its estimate) ^ ^ ^

  23. Distribution of µ used for CI - relates to “range” for mean Fitted distribution Prediction interval for µ unknown – relates to “range” for future observed So again, imagine that the distributions are normal. The next observation might lie – down here, or up here. (implications for risk capital) ^

  24. CI Returning to example: 95% prediction interval for Yn+1is (6.75,7.63):

  25. Parameter uncertainty can be reduced - more data reduces parameter uncertainty (more than 10 tosses of the coin in pre-trial). In some cases you can go back and get more loss data (e.g. dig up an older year) – or eventually you'll have another year of data • But process variability doesn't reduce with more data - an aspect of the process, not knowledge

  26. Note that nearby developments are related: • If DY 3 was all missing, you could take a fair guess at where it was! • So in this case we do have much more data!

  27. Log-Normalized vs Dev Year 8 7.8 7.6 7.4 7.2 7 6.8 6.6 6.4 6.2 6 CI PI 2 3 4 • Need a model to take full advantage of this. Even just fitting line through DY2-4 has a big effect on the width of the confidence interval: • Only changes the prediction interval by ~2%. So calculated V@R hardly changes

  28. But that prediction interval is on log scale. • To take a prediction interval back to the normalized-$ scale, just back-transform the endpoints of the PI • To produce a confidence interval for the mean on the normalized-$ scale is harder (can’t just backtransform limits on CI – that’s an interval for the median) • not a particularly enlightening bit of algebra, so we’re leaving the derivation out here

  29. There are some companies around for whom (for some lines of business) the process variance is very large. - some have a coefficient of variation near 0.6. [so standard deviation is > 60% of the mean] That's just a feature of the data. - May not be able to control it, but you sure need to know it.

  30. Why take logs? • tends to stabilize variance • multiplicative effects (including economic effects, such as inflation) become additive (percentage increases or decreases are multiplicative) - exponential growth or decay → linear - skewness often eliminated distributions tend to look near normal - Using logs a familiar way of dealing with many of these issues (standard in finance) NB for these to work have to take logs of incremental paid, not cumulative paid.

  31. x x x x x x x x x x x x x x x x x x x x x x x x x x ProbabilisticModelling of trends e.g. trends in the development year direction If we graph the data for an accident year against development year, we can see two trends. 0123456789101112

  32. x x x x x x x x x x x x x x x x x x x x x x x x x x Probabilistic Modelling Could put a line through the points, using a ruler. Or could do something formally, using regression. 0123456789101112 Variance =

  33. Introduction to Probabilistic Modelling Models Include More Than The Trends x x x (y – ŷ) x x x x x x x x x x 0 1 2 3 4 5 6 7 8 9 10 11 12 • The model is not just the trends in the mean, but the distribution about the mean (Data = Trends + Random Fluctuations)

  34. o x o o o x x x x o x o o x x o o x o o x o x x x o Introduction to Probabilistic Modelling Simulating the Same “Features” in the Data 0 1 2 3 4 5 6 7 8 9 10 11 12 • Simulate “new” observations based in the trends and standard errors • Simulated data should be indistinguishable from the real data

  35. - Real Sample: x1,…,xn - Random Sample from fitted distribution: y1,…,yn What does it mean to say a model gives a good fit? e.g. lognormal fit to claim size distribution Does not mean we think the model generated the data fitted lognormal - Fitted Distribution y’s look like x’s: — Model has probabilistic mechanisms that can reproduce the data

  36. PROBABILISTIC MODEL S3 Trends+ variationabout trends Real Data S2 S1 Based on Ratios Simulated triangles cannot be distinguished from real data – similar trends, trend changes in same periods, same amount of random variation about trends Models project past volatility into the future

  37. Trends in three directions, plus volatility “picture” of the model

  38. Testing a ratio model • Since the correctness of an interval (“range”) depends on the model, it’s necessary to test a model’s appropriateness • Simple diagnostic model checks exist • Ratio models can be embedded in regression models, and so we can do more rigorous testing – extend this to a diagnostic model

  39. j j-1 y X = Cum. @ j-1 Y = Cum. @ j y y x x x ELRF(Extended Link Ratio Family)x is cumu. at dev. j-1 and y is cum. at dev. j • Link Ratios are a comparison of columns • We can graph the ratios of Y:X - line through O? y/x y/x Using ratios => E(Y|x) = x

  40. Mack (1993) Chain Ladder Ratio( Volume Weighted Average) Arithmetic Average

  41. Intercept (Murphy (1994)) Since y already includes x: y = x + p Incremental Cumulative at j at j -1 Is b -1 significant ? Venter (1996)

  42. j-1 j Cumulative Incremental p j-1 j } p x x x x x x Use link-ratios for projection x Abandon Ratios - No predictive power

  43. p x Is assumption E(p|x) = a + (b-1) xtenable? Note: If corr(x, p) = 0, then corr((b-1)x, p) = 0 If x, p uncorrelated, no ratio has predictive power Ratio selection by actuarial judgement can’t overcome zero correlation. • Corr. often close to 0 • Sometimes not. • Does this imply ratios • are a good model? • - ranges?

  44. X Y N X Y • With two associated variables, tempting to think X causes changes in Y. • However, may have another variable impacting both X and Y in a similar way, causing them move together: Here X is a noisy proxy for N; N is a better predictor of Y (sup. inflation, exposures)

  45. Cumulative Incremental j-1 j } p j-1 j 90 91 92 w Condition 1: p x x x x x x x x w p Condition 2:

  46. Now Introduce Trend Parameter For Incrementals p p vs acci. yr, not previous cumulative NB: diagnostic model, not predictive model

  47. Condition 3: Incremental Review 3 conditions: Condition 1: Zero trend Condition 2: Constant trend, positive or negative Condition 3: Non-constant trend

  48. Past Future 1986 1987 1998 Probabilistic Modelling Trends occur in three directions: 0 1 Development year d Calendar year t = w+d w Accident year

  49. - 0.2d d M3IR5 Data 0 1 2 3 4 5 6 7 8 9 10 11 12 13 100000 81873 67032 54881 44933 36788 30119 24660 20190 16530 13534 11080 9072 7427 100000 81873 67032 54881 44933 36788 30119 24660 20190 16530 13534 11080 9072 100000 81873 67032 54881 44933 36788 30119 24660 20190 16530 13534 11080 100000 81873 67032 54881 44933 36788 30119 24660 20190 16530 13534 100000 81873 67032 54881 44933 36788 30119 24660 20190 16530 100000 81873 67032 54881 44933 36788 30119 24660 20190 100000 81873 67032 54881 44933 36788 30119 24660 100000 81873 67032 54881 44933 36788 30119 100000 81873 67032 54881 44933 36788 100000 81873 67032 54881 44933 100000 81873 67032 54881 100000 81873 67032 100000 81873 100000 alpha = 11.513 -0.2 PAID LOSS = EXP(alpha - 0.2d)

  50. 0.15 0.3 0.1 Probabilistic Modelling Axiomatic Properties of Trends

More Related